Efficient visualization of large medical image datasets on standard PC hardware

Fast and accurate algorithms for medical image processing and visualization are becoming increasingly important due to routine acquisition and processing of rapidly growing amounts of data in clinical practice. At the same time, standard computer hardware is becoming sufficiently powerful to be used in applications which previously required expensive and inflexible special-purpose hardware. We present an efficient volume rendering approach using the example of maximum intensity projection (MIP), which is an important clinical tool. The method systematically exploits the properties of general-purpose hardware such as hierarchical cache memories and super-scalar processing. In order to optimize the cache efficiency, the dataset is processed in blocks which fit into the processor cache. The innermost ray casting loop is transformed such that the arithmetic operations and memory accesses can be processed in parallel on current general-purpose processors. Combined with other optimization strategies, such as vectorization and block-wise ray skipping, this approach yields near-interactive frame rates for large clinical datasets using a standard dual-processor PC. Data compression and simplification methods have intentionally not been used in order to demonstrate the achievable performance without any quality reductions. Some of the presented ideas can be applied to other computationally intensive image processing tasks.